CUDA position independent code

[rijobro]

CUDA parts of NiftyReg are static libraries. In SIRF, some libraries that link against these CUDA NiftyReg libraries need to be static. This causes problems, as per issue https://github.com/CCPPETMR/SIRF/issues/551.

The solution is to add the following to the NVCC compiler options: --compiler-options -fPIC.

Would you have any problem with me making a PR to add this in? Should it be added by default, or would you like me to create an option, which when set to ON, appends the above string? It should be a 2-minute job, so let me know what you prefer.

[rijobro]

proposition 1:

set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --compiler-options -fPIC")

proposition 2:

if (CMAKE_POSITION_INDEPENDENT_CODE)
    set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --compiler-options -fPIC")
endif()

proposition 3:

if (CMAKE_POSITION_INDEPENDENT_CODE)
    if("${CMAKE_SYSTEM}" MATCHES "Linux")
        set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --compiler-options -fPIC")
    endif()
endif()

[KrisThielemans]

All of these are dangerous as each compiler might need a different flag. proposition 4:

 if(CMAKE_POSITION_INDEPENDENT_CODE AND DEFINED CMAKE_C_COMPILE_OPTIONS_PIC)
       # add (undocumented) CMake flag that should tell the host compiler to generate position independent code
        set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --compiler-options ${CMAKE_C_COMPILE_OPTIONS_PIC}")
endif()

[KrisThielemans]

I wouldn't do this by default for sure. position-independent code can be less efficient on some architectures.

I see no need for a new option. CMake has CMAKE_POSITION_INDEPENDENT_CODE already. If anyone wnats to build NiftyReg with this, they can just set CMAKE_POSITION_INDEPENDENT_CODE=ON